Are Data Centres Really Why Your Power Bill Went Up?

Ask why electricity got more expensive and you will get two answers delivered with equal certainty. One says AI data centres are draining the grid and you are paying for it. The other says data centres are a convenient scapegoat and the real culprit is bad state energy policy. Both camps have numbers. Both are leaving things out.

The case that data centres are raising bills

This one has the strongest evidence, and it is regional. In the PJM territory, which covers thirteen states and Washington DC, the capacity auction for 2026 to 2027 cleared at 329.17 dollars per megawatt-day, a record, and above the cap the Federal Energy Regulatory Commission had approved. Ratepayers across the region will pay roughly 1.4 billion dollars more in capacity costs starting this June.

The attribution is not guesswork either. Analysis from IEEFA attributed 63 percent of the increase in the previous auction to data centre demand, around 9.3 billion dollars in added capacity costs spread across all customers. Over four auctions the running total is about 29.4 billion. The Citizens Utility Board renewed its call for reform on 15 July on exactly these grounds.

So in PJM, if you are a household paying more, some of that is data centre demand arriving faster than new generation. That is not a talking point. It is an auction result.

The case that they are not

Now the other side, which is doing something subtler than it looks. Writers at RealClearEnergy and elsewhere compared residential rates against data centre concentration across all fifty states and found the correlation weak. The steepest increases have come from West Coast states with relatively few data centres, while Virginia, where data centres take roughly a quarter of electricity consumption, sits at 15.94 cents per kilowatt hour against a national average of 17.24.

That finding is real and it is worth sitting with. A big industrial customer contributes to fixed costs that get spread across everyone, so extra load can push average rates down as easily as up, depending on whether the utility had spare capacity.

What the argument skates over is timing and place. A national scatter plot averages away the specific thing that hurts, which is fast load growth arriving in a constrained region before new supply is built. Virginia's rates reflect decades of buildout. PJM's auction reflects the last eighteen months. Both can be true.

It is also worth noticing that much of this analysis comes from outlets with a settled position on fossil fuel policy, and the conclusion arrives conveniently pre-shaped. That does not make the data wrong. It does mean you should check what was measured.

What the global numbers say

Zooming out flattens the drama considerably. The International Energy Agency puts data centres at about 415 terawatt hours in 2024, roughly 1.5 percent of world electricity, rising to around 945 terawatt hours by 2030. That is a doubling, and it is still just under 3 percent of the total. Data centres account for less than a tenth of global electricity demand growth this decade.

The IEA adds the qualifier that matters more than the headline: unlike electric vehicles, data centres cluster. Diffuse demand growth is easy for a grid to absorb. Concentrated demand growth is what breaks things, and it breaks them locally.

Where this leaves you

The claim that AI is why your bill went up is wrong as a general statement about your country and defensible as a statement about specific grids in specific years. The claim that data centres have nothing to do with it requires ignoring an auction that priced their demand explicitly.

If you want to know whether AI is on your bill, the useful question is not what AI uses globally. It is whether your utility recently added a large load faster than it added supply. In much of PJM the answer is yes. In much of the country it is no, and your rates went up for older and duller reasons.

We have written before about the claim that one AI prompt drinks a bottle of water, which fails in a similar way. The resource cost is real, the aggregate is smaller than the viral version, and the local impact is what people actually experience.